0
$\begingroup$

Do outliers change the outcome of time series decomposition?

As far as I understand it, outliers occur in the residual-component. In the residuals plot they can be visually identified as spikes.

However, if I know the outliers within a time series before applying the decomposition, does it make sense to remove the outliers in order to increase the "quality" of the decomposition?

$\endgroup$
1
  • 1
    $\begingroup$ You may want outliers to show mostly in residuals, but that hope or expectation may not be satisfied, depending largely on how your decomposition is fitted, and whether the fitting method is robust. Ignoring outliers is usually regarded as a bad idea. In the case of time series there is a real problem that you have to fill gaps for many calculations to make sense at all. $\endgroup$
    – Nick Cox
    Commented Jun 29, 2020 at 14:17

2 Answers 2

1
$\begingroup$

See section 12.9 of Hyndman and Athanasopoulos's textbook. https://otexts.com/fpp2/missing-outliers.html

Often in a time series you need to handle outliers -- e.g. if you are predicting airline passengers you will have big outliers around Sept 11, 2001 and currently due to COVID-19. These are special events that have to be handled or they will throw everything off. So, yes, you need to handle these before the decomposition.

Data errors, or likely data errors, have to go.

Otherwise be careful. Be particularly careful in just applying some global filter (e.g. everything outside 3 standard deviations). You may be just suppressing the naturally high variability of the data.

$\endgroup$
1
$\begingroup$

In the cross sectional literature it is generally seen as a bad idea to remove an outlier not caused by a data error. Instead you should try to understand why it is occurring. Which explains your data better. I have not seen it addressed much in time series.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.